Split tofu-risk TS character mappings into an extension table

[frankslin]

Keep tofu-risk mappings in TSCharacters.txt as non-default candidates by making the source character the first candidate on annotated rows. This removes those rare inferred simplified forms from the default TSCharacters conversion result when the extension table is not loaded.

Update extract_tofu_risk.py so TSCharactersExt.txt is generated from the same annotations while skipping the newly inserted identity candidate. The generated extension table therefore preserves the old risk mappings, allowing configurations that load it before TSCharacters.ocd2 to keep existing conversion behavior.

Wire TSCharactersExt into Bazel, CMake, Node gyp, npm packaging, and the t2s/hk2s/tw2s/tw2sp configs, including the Jieba tw2sp config. The extension dictionary is generated and compiled before TSCharacters.ocd2 in those configurations so current default software behavior remains unchanged.

Verified with: bazel build //data/dictionary:generate_bin_TSCharactersExt; bazel test //data/config:config_dict_validation_test; bazel test //test:command_line_converter_test; bazel test //data/dictionary:dictionary_test; bazel test //data/config:config_schema_validation_test; cmake --build build/dbg --target Dictionaries; ./node_modules/.bin/node-gyp rebuild; node scripts/prepare-node-prebuild-artifacts.js; python3 -m build --wheel.

Note: npm test was also run, but it still fails existing fruit-drying testcase expectations unrelated to TSCharactersExt packaging.

[frankslin]

@danny0838 这个版本可以测试了，目前的行为保持和原来一样。把几个 json 里的 TSCharactersExt.ocd2 字典去掉，就变成不生成「豆腐块」的新行为了。

Ref: #217